CLIMB-3D: Continual Learning for Imbalanced 3D Instance Segmentation

BMVC 2025

Vishal Thengane, Jean Lahoud, Hisham Cholakkal, Rao Muhammad Anwer, Lu Yin, Xiatian Zhu, Salman Khan

University of Surrey, University of Wollongong

vgthengane (at) gmail (dot) com

Paper

ArXiv Code

CLIMB-3D: Continual Learning for Imbalanced 3D Instance Segmentation teaser

Continual Learning 3D Vision Instance Segmentation

Abstract

CLIMB-3D introduces class-incremental learning for point cloud instance segmentation and evaluates it on long-tail settings derived from ScanNet200. The method addresses class imbalance to improve performance across both frequent and rare classes.

Method

The framework formulates continual point cloud instance segmentation with class imbalance in mind. It uses an incremental training strategy with imbalance-aware components to preserve prior knowledge while learning new classes more uniformly.

Results

Across long-tail splits derived from ScanNet200, CLIMB-3D improves performance consistency between frequent and rare classes while maintaining strong overall segmentation quality in incremental settings.

BibTeX

@inproceedings{thengane2025climb3d,
  title={CLIMB-3D: Continual Learning for Imbalanced 3D Instance Segmentation},
  author={Thengane, Vishal and Lahoud, Jean and Cholakkal, Hisham and Anwer, Rao Muhammad and Yin, Lu and Zhu, Xiatian and Khan, Salman},
  booktitle={BMVC},
  year={2025}
}